[llvm-dev] [LICM][MemorySSA] Converting LICM pass to use MemorySSA to avoid AliasSet collapse issue

Geoff Berry via llvm-dev llvm-dev at lists.llvm.org
Wed Apr 20 12:06:50 PDT 2016

1)      Sounds good.  This isn’t holding me up so I’ll just try to keep an eye out for these changes.


2)      I’ve attached an example IR file and debug log of where the caching is going bad.  It depends on my changes to EarlyCSE, but hopefully it is clear from the debug output what is going on.  Let me know if there is a better way to get this repro case to you.  Also, I’ll be on IRC for the next couple of hours if you would like to have a quicker discussion.



Geoff Berry

Employee of Qualcomm Innovation Center, Inc.

Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum, a Linux Foundation Collaborative Project


From: Daniel Berlin [mailto:dberlin at dberlin.org] 
Sent: Wednesday, April 20, 2016 1:06 PM
To: Geoff Berry <gberry at codeaurora.org>; George Burgess <gbiv at google.com>
Cc: llvm-dev <llvm-dev at lists.llvm.org>
Subject: Re: [LICM][MemorySSA] Converting LICM pass to use MemorySSA to avoid AliasSet collapse issue




On Wed, Apr 20, 2016 at 9:58 AM, Geoff Berry <gberry at codeaurora.org <mailto:gberry at codeaurora.org> > wrote:

Hi Daniel,


Thanks for the info.  I’ve started looking into converting EarlyCSE to use MemorySSA first since 1) I don’t think it needs any additional MemorySSA update API and 2) the particular case I’m looking at needs EarlyCSE to catch more load cases before LICM to be profitable.

I have a prototype working, but have run into two issues:


1)      readonly calls are treated as clobbers by MemorySSA which leads to extra walking of MemoryDefs to not regress some EarlyCSE test cases.  This isn’t a huge deal, I’m just wondering if it is intentional or something that just hasn’t been gotten to yet.


George is working on the optimizations, of which this is one.

I think this is one of the ones his current patch (under review) addresses.


2)      There seems to be a bug in the CachingMemorySSAWalker invalidation causing it to return MemoryAccess nodes that have been removed.  In the case I’m seeing, a call node is removed from MemorySSA which causes CachingMemorySSAWalker::invalidateInfo() to clear the CachedUpwardsClobberingCall map.  However, this same call node is present as a value in the CachedUpwardsClobberingAccess map,


Unless i'm missing something, this should not have happened, and we should assert they are not being added to the cache.


The truth is the caching parts are complicated and ugly. It was meant to be a pretty simple cache, but it's known to be inefficient (memory wise) and it's on the list of things to clean up and make sane.


Do you have a testcase where this happens?

A quick glance says we check whether it's a call in all the right places, which means there must be a place we are not *setting* isCall properly.


-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20160420/f04e9baf/attachment.html>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: memssa-remove-cache-bug.ll
Type: application/octet-stream
Size: 263 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20160420/f04e9baf/attachment.obj>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: memssa-remove-cache-bug.log
Type: application/octet-stream
Size: 2726 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20160420/f04e9baf/attachment-0001.obj>

More information about the llvm-dev mailing list